Using Large Languge Models for Processing Sensor Data

Area of Science:

• Data Science
• Artificial Intelligence
• Sensor Technology

Background:

• Sensor data is often stored in varied, inaccessible formats, hindering its reuse.
• Extracting meaningful information from unstructured and semi-structured text containing sensor data presents a significant challenge.

Purpose of the Study:

• To develop and evaluate a workflow for extracting sensor data from text using Large Language Models (LLMs).
• To compare the performance of different LLMs (GPT-4, Llama 3, Mistral, Falcon) in sensor data extraction and structuring.
• To establish standardized JSON output for automated data processing.

Main Methods:

• Utilized Large Language Models including GPT-4, Llama 3, Mistral, and Falcon for data extraction.
• Employed careful prompt engineering to enforce a strict JSON output structure.
• Developed new metrics for comparing the efficiency and accuracy of different models.
• Tested model performance on both freeform and tabular text data.

Main Results:

• GPT-4 demonstrated high conversion efficiency, while open-source models showed comparable performance with appropriate data structures.
• Smaller models struggled with freeform text but excelled with tabular data.
• Larger models proved more robust in handling freeform text and minimizing conversion errors.
• A multi-purpose workflow for sensor data extraction was successfully established.

Conclusions:

• LLMs, particularly larger ones, offer a viable solution for standardizing diverse sensor data formats through structured extraction.
• Prompt engineering and appropriate data structuring are crucial for optimizing LLM performance in sensor data extraction tasks.
• The developed workflow facilitates automated processing and enhances the reusability of sensor data from various text sources.